FlashMap: flow map-based speculative decoding speeds up Qwen3-8B inference over 6x
caglarml · x · 2026-10-25
- FlashMap is a new speculative decoding algorithm built on flow map language models, released by jeffshen and collaborators with a project page and arXiv paper.
- The key insight: flow map LMs capture token dependencies in just 1-2 forward passes, making them ideal drafters for speculative decoding — achieving over 6x faster inference on Qwen3-8B.
- sedielem praised the results and highlighted the shuffle test used to validate the approach.
More from Models
- Ternary weights shrink a speech model from 2,472 MB to 181 MB with just +0.4pt WER — lmoroney · 2026-10-25
- Unverified DeepSeek-V4.1-Flash appears on Hugging Face with 763B params, llama.cpp PR filed — jacek2023 · 2026-10-25
- DeepSeek-V4.1-Flash EXL3-3bpw quant runs on 2x RTX PRO 6000 — No-Paper-557 · 2026-10-25
- Dev argues Codex still beats Opus 5.5 at coding: Opus brute-forces bad designs instead of researching — haider1 · 2026-10-25
- Claude Opus now refuses to decompile code, while GPT-6 Astra does it happily — CtrlAltDwayne · 2026-10-25
- User says Hermes feels like an infinite-level game, 'brilliant and scary' as it compounds — Teknium · 2026-10-25